From e75e08864bbf0dc7dca7783e8d2a1fb77a7a1b39 Mon Sep 17 00:00:00 2001 From: Sameer Kankute Date: Fri, 1 May 2026 18:36:51 +0530 Subject: [PATCH] fix(xai): normalize usage total_tokens for prompt caching xAI can return total_tokens inconsistent with prompt_tokens + completion_tokens when caching is enabled. Align with OpenAI-style usage so shared LLM tests and downstream consumers see coherent totals. Apply to non-streaming responses and streaming usage chunks. Made-with: Cursor --- litellm/llms/xai/chat/transformation.py | 28 ++++++++++++++++++++++++- 1 file changed, 27 insertions(+), 1 deletion(-) diff --git a/litellm/llms/xai/chat/transformation.py b/litellm/llms/xai/chat/transformation.py index 6300868a641..ccb81840e59 100644 --- a/litellm/llms/xai/chat/transformation.py +++ b/litellm/llms/xai/chat/transformation.py @@ -1,4 +1,4 @@ -from typing import Any, AsyncIterator, Iterator, List, Optional, Tuple, Union +from typing import Any, AsyncIterator, Dict, Iterator, List, Optional, Tuple, Union import httpx @@ -26,6 +26,7 @@ from ...openai.chat.gpt_transformation import ( class XAIChatConfig(OpenAIGPTConfig): + @property def custom_llm_provider(self) -> Optional[str]: return "xai" @@ -225,6 +226,9 @@ class XAIChatConfig(OpenAIGPTConfig): verbose_logger.debug(f"Error extracting X.AI web search usage: {e}") self._fold_reasoning_tokens_into_completion(response) + self._normalize_openai_compatible_usage_totals( + getattr(response, "usage", None) + ) return response @staticmethod @@ -284,6 +288,25 @@ class XAIChatConfig(OpenAIGPTConfig): setattr(usage, "num_sources_used", int(num_sources_used)) verbose_logger.debug(f"X.AI web search sources used: {num_sources_used}") + @staticmethod + def _normalize_openai_compatible_usage_totals( + usage: Union[Usage, Dict[str, Any], None], + ) -> None: + if usage is None: + return + if isinstance(usage, dict): + prompt_tokens = int(usage.get("prompt_tokens") or 0) + completion_tokens = int(usage.get("completion_tokens") or 0) + expected_total = prompt_tokens + completion_tokens + if int(usage.get("total_tokens") or 0) != expected_total: + usage["total_tokens"] = expected_total + return + prompt_tokens = int(usage.prompt_tokens or 0) + completion_tokens = int(usage.completion_tokens or 0) + expected_total = prompt_tokens + completion_tokens + if int(usage.total_tokens or 0) != expected_total: + usage.total_tokens = expected_total + class XAIChatCompletionStreamingHandler(OpenAIChatCompletionStreamingHandler): def chunk_parser(self, chunk: dict) -> ModelResponseStream: @@ -304,4 +327,7 @@ class XAIChatCompletionStreamingHandler(OpenAIChatCompletionStreamingHandler): # Add a dummy choice with empty delta to ensure proper processing chunk["choices"] = [{"index": 0, "delta": {}, "finish_reason": None}] + if "usage" in chunk and chunk["usage"] is not None: + XAIChatConfig._normalize_openai_compatible_usage_totals(chunk["usage"]) + return super().chunk_parser(chunk)